We present an assessment of the accuracy of common operations performed in 21-cm spectral line stacking experiments. To this end, we generate mock interferometric data surveying the 21-cm emission at frequency 1310 < ν < 1420 MHz (0.005 < z < 0.084) and covering an area ~6 deg2 of the sky, mimicking the observational characteristics of real MeerKAT observations. We find that the primary beam (PB) correction accounts for just few per cent ($\sim 8{{\ \rm per\ cent}}$ at 0 PB power, $\sim 3{{\ \rm per\ cent}}$ at 0.6 PB power) deviations from the true MHI signal, and that weighting schemes are based on noise properties provide unbiased results. On the contrary, weighting schemes based on distance can account for significant systematic mass differences when applied to a flux-limited sample ($\Delta M_{\rm HI}\sim 40{-}50{{\ \rm per\ cent}}$ in the studied case). We find no significant difference in the final $\mathinner {\langle {M_{\rm HI}}\rangle }$ obtained when spectroscopic redshift uncertainties are accounted for in the stacking procedure (Δz ~ 0.000 35, i.e. $\Delta v \sim 100\, {\rm km\, s}^{-1}$). We also present a novel technique to increase the effective size of the galaxy sample by exploiting the geometric symmetries of galaxy cubelets, potentially enhancing the S/N by a factor of $\sim \sqrt{2}$ when analyzing the final stacked spectrum (a factor of 4 in a cubelet). This procedure is found to be robustly unbiased, while efficiently increasing the S/N, as expected. We argue that an appropriate framework employing detailed and realistic simulations is required to exploit upcoming data sets from SKA pathfinders in an accurate and reliable manner.